Hardip Bhesaniya

Position stated plainly

This is independent work. I hold no institutional affiliation, no academic supervision and no research funding for it, and this page claims no degree, title or endorsement. Everything below describes what I have built and how I work — which is the part that can be checked.

I would rather be assessed on the software and its verification record than on credentials. The case study is written so that a specialist can attempt to falsify it in about a minute, and the package ships the fixtures and hashes to do so.

What I am responsible for

Numerical kernels

The shared calculation modules behind all seven laboratories: the Maxwellian rate integral, relativistic invariant-s reaction kinematics, analytic mirror and toroidal field expressions, the fixed-boundary Grad–Shafranov solver, relativistic Boris orbit integration, the image segmentation pipeline and the electrostatic conversion ledger.

Data provenance

Selecting, transcribing and hashing the bundled datasets — the Sikora–Weller cross-section evaluation, the AME2020 mass slice, Bosch–Hale reactivity fits, IAEA LiveChart ground-state records — and making sure every result can be traced back to the exact bytes it was computed from.

Verification

Writing the benchmark suite and its acceptance tolerances, building independent reference implementations to compare against, and keeping the distinction between an internal consistency check and an external validation explicit in the record.

Claim discipline

Deciding what the software is allowed to say. Every model contract names its equations, its validity domain and the mechanisms it omits, and a result outside its fit domain stays visible and flagged rather than being quietly clipped.

Release integrity

The manifest, checksums and source inventory; making sure the shipped package is the package the benchmarks ran against, and that anyone can verify that themselves.

Interface and delivery

The full front end: the seven laboratory workspaces, the three-dimensional scenes with a CPU fallback path for machines without WebGL, and the static hosting package that runs with no server-side service and no network dependency.

Methods I work in

• Reaction rate integrationMaxwellian averaging of a tabulated cross section with log–log interpolation and segment boundaries at every tabulated energy; exact invariant-s transformation between laboratory and centre-of-mass axes; quadrature verified against a synthetic constant-cross-section oracle to a relative error of 1.6 × 10⁻¹².
• Relativistic kinematicsQ values and thresholds from bare-nucleus masses, exact two-body partitions, and a sequential three-body decay sampler that enforces energy, momentum and mass-shell closure event by event under fixed seeds.
• Charged-particle orbitsRelativistic Boris integration in prescribed analytic fields, verified by kinetic-energy invariance to 3.2 × 10⁻¹⁴ over 100 gyroperiods and by second-order timestep refinement against an integrated analytic orbit.
• Elliptic solversSecond-order centred-difference discretisation of the Grad–Shafranov operator with the cylindrical term carried explicitly, verified by manufactured solution at an observed order of 2.0020 and cross-checked node-for-node against an independent Python implementation.
• Image analysisOtsu and manual thresholding, 4- and 8-connected component labelling, per-object morphometry, and Poisson counting intervals using exact central Garwood below n = 100 with an identified Wilson–Hilferty approximation above.
• Numerical software practiceTolerances declared before a comparison runs; independent implementations rather than self-comparison; deterministic seeds; hashed source inventories attached to every benchmark record; and reproducibility tested across engine versions rather than assumed.

How I work

Three habits shape the whole package, and they are the ones worth judging.

Tolerances come first. Every benchmark fixes its acceptance criterion before the comparison runs. A tolerance chosen after seeing the answer is not a test.

Independence over agreement. Where a result matters, it is checked against something written separately — a closed form, a published anchor, or a program in another language that never reads the output it is checking. Two implementations agreeing is weak evidence if one was written by looking at the other.

The omission list is part of the result. Each model contract states what it does not include. A balance that silently drops synchrotron radiation, impurities or transport is not a smaller model, it is a wrong one, and saying so is cheaper than being found out.